Clustering Methods in Symbolic Data Analysis

We present an overview of the clustering methods developed in Symbolic Data Analysis to partition a set of conceptual data into a fixed number of classes. The proposed algorithms are based on a generalization of the classical Dynamical Clustering Algorithm (DCA) (Nuées Dynamiques méthode). The criterion optimized in DCA is a measure of the fit between the partition and the representation of the classes. The prototype is a model of the representation of a class, and it can be an element of the same space of representation as the symbolic data to be clustered or, according to the nature of the data, the prototype is a higher-order object which generalizes the characteristics of the elements belonging to the class. The allocation function for the assignment of the objects to the classes depends on the nature of the variables which describe the symbolic objects. The choice of such function must be related to the particular type of prototype preferred as the representation model of the classes.